Mycobacterium tuberculosis and Clostridium difficille interactomes: demonstration of rapid development of computational system for bacterial interactome prediction
- Seshan Ananthasubramanian1, 2,
- Rahul Metri3,
- Ankur Khetan†4,
- Aman Gupta†5,
- Adam Handen†6,
- Nagasuma Chandra3 and
- Madhavi Ganapathiraju1, 2Email author
© Ananthasubramanian et al; licensee BioMed Central Ltd. 2012
Received: 13 October 2011
Accepted: 21 March 2012
Published: 21 March 2012
Protein-protein interaction (PPI) networks (interactomes) of most organisms, except for some model organisms, are largely unknown. Experimental methods including high-throughput techniques are highly resource intensive. Therefore, computational discovery of PPIs can accelerate biological discovery by presenting "most-promising" pairs of proteins that are likely to interact. For many bacteria, genome sequence, and thereby genomic context of proteomes, is readily available; additionally, for some of these proteomes, localization and functional annotations are also available, but interactomes are not available. We present here a method for rapid development of computational system to predict interactome of bacterial proteomes. While other studies have presented methods to transfer interologs across species, here, we propose transfer of computational models to benefit from cross-species annotations, thereby predicting many more novel interactions even in the absence of interologs. Mycobacterium tuberculosis (Mtb) and Clostridium difficile (CD) have been used to demonstrate the work.
We developed a random forest classifier over features derived from Gene Ontology annotations and genetic context scores provided by STRING database for predicting Mtb and CD interactions independently. The Mtb classifier gave a precision of 94% and a recall of 23% on a held out test set. The Mtb model was then run on all the 8 million protein pairs of the Mtb proteome, resulting in 708 new interactions (at 94% expected precision) or 1,595 new interactions at 80% expected precision. The CD classifier gave a precision of 90% and a recall of 16% on a held out test set. The CD model was run on all the 8 million protein pairs of the CD proteome, resulting in 143 new interactions (at 90% expected precision) or 580 new interactions (at 80% expected precision). We also compared the overlap of predictions of our method with STRING database interactions for CD and Mtb and also with interactions identified recently by a bacterial 2-hybrid system for Mtb. To demonstrate the utility of transfer of computational models, we made use of the developed Mtb model and used it to predict CD protein-pairs. The cross species model thus developed yielded a precision of 88% at a recall of 8%. To demonstrate transfer of features from other organisms in the absence of feature-based and interaction-based information, we transferred missing feature values from Mtb orthologs into the CD data. In transferring this data from orthologs (not interologs), we showed that a large number of interactions can be predicted.
Rapid discovery of (partial) bacterial interactome can be made by using existing set of GO and STRING features associated with the organisms. We can make use of cross-species interactome development, when there are not even sufficient known interactions to develop a computational prediction system. Computational model of well-studied organism(s) can be employed to make the initial interactome prediction for the target organism. We have also demonstrated successfully, that annotations can be transferred from orthologs in well-studied organisms enabling accurate predictions for organisms with no annotations. These approaches can serve as building blocks to address the challenges associated with feature coverage, missing interactions towards rapid interactome discovery for bacterial organisms.
The presence of about 500-1,000 bacterial species in the human gut flora of the intestines plays important role in immunity and nutrition . While some bacteria live in a symbiotic relationship with humans, numerous others cause diseases, killing millions of people annually. Mycobacterium tuberculosis (Mtb) causes tuberculosis which remains a leading infectious disease to this day, with about 2 million deaths annually worldwide [2–4]. A deadly synergy with human immunodeficiency virus (HIV) further increases the burden of the disease [5, 6]. Clostridium difficile (CD) infection is the primary cause of antibiotic-associated diarrhoea. It has a property of undergoing mutation rapidly . In the past ten years, variant toxin-producing strains of C. difficile have emerged, that have been associated with severe disease outbreaks worldwide . Understanding the functions of the proteins of the pathogens, and their interactions with each other and with host proteins would provide a basis for understanding the pathogenesis and virulence mechanisms of the pathogen [9–11]. Analyses of protein-protein interaction (PPI) networks have led to rich insights about possible pathways of information flow that could lead to the emergence of drug resistance in the pathogen . A better understanding of pathogen-interactome can also help in studying host-pathogen interactions [13, 14]. Despite the importance of PPIs, the high cost, time-consuming and labor intensive nature of experimental methods have resulted in limited availability of currently known PPIs.
Newer techniques such as mass-spectrometry, yeast two-hybrid (Y2H), and tandem affinity purification have been designed to study the global interactions in an organism in a high-throughput manner. Y2H methods have a low recall, in the order of about 12% . This low recall, coupled with the fact that only one in a thousand pairs of proteins is an interacting pair, makes it expensive as well as infeasible to characterize whole interactomes with only high throughput biotechnology. Data explosion in the post-genomic era has resulted in advancement in computational methods to extract and analyze useful biological information, including PPIs. Computational methods for PPI prediction use biological information such as genomic data, 3D structures, gene co-expression, co-occurrence, co-evolution, and co-localization as features based on which a model is learnt to classify interacting pairs from non-interacting pairs . Although computational methods are being developed for discovering more and more new PPIs in organisms, the required feature-base (gene expression profiles, gene ontology annotations, etc) is not available except for a few well studied organisms such as yeast, mouse and human. For other organisms for which a rich feature-base is not available, some PPIs may be inferred bioinformatically by identifying interactions among orthologs (namely, by identifying interologs) [17, 18] or with computational methods that use only protein sequence information. Phylogenetic profiling infers protein interactions from patterns of presence or absence of proteins across multiple genomes. If genes are functionally related in an organism, co-inheritance of the genes would occur in other organisms, since the absence of any of the genes would result in the loss of function of the other genes . Gene fusion predicts interactions between two proteins in an organism based on the evidence that they form a part of a single protein in a different genome . Gene order conservation infers interactions based conserved neighbourhood of genes on the genome . Similarity of Phylogenetic trees assumes that interacting protein sequences and their partners must co-evolve and pairs of protein sequences exhibiting high degree of co-evolution are inferred to be interacting . Some methods adopt sequence-based approaches, which include sequence co-evolution, a method based on multiple sequence alignments and domain pairs-based approaches, where domain-domain interactions are noted to be conserved across species . Homology-based approaches are based on both sequence and structure information to transfer known interactions [23, 24]. Some studies have used model organisms to aid in the discovery of the interactome of another organism. Using domain-domain interactions, it has been shown that it may be significantly more precise to predict PPIs from multiple organisms than from a single organism . Streptococcus pneumoniae interactome determined with Y2H was compared with that of E. coli, and a large number of conserved interactions were found among the orthologs between the two organisms .
Results and discussion
Predicting PPIs is treated as a classification problem in machine learning. Each protein-pair (an instance in unlabelled data) is classified as interacting or non-interacting based on the features describing that pair. A classifier (e.g. a random forest or a support vector machine) is learned from the features of the instances whose interactions are known (labelled data). We use these STRING and GO annotations in building features of protein-pairs. For labels, instances that have nonzero value for STRING experimental score are treated as known interactions. Simply sequencing the genome of a bacterium provides genomic context features in STRING. Several bacteria have at least partial annotations available in GO. Of the 1,707 genomes represented in GO, 1,194 have 50-70% of their genomes annotated and only 98 genomes have less than 50% annotated .
1-(888 + 868)/(4020 + 4147) = 0.785
1-(735 + 523)/(6717 + 4020) = 0.883
1-(472 + 452)/(22500 + 4020) = 0.9652
1-(374 + 603)/(6717 + 4147) = 0.91
1-(366 + 338)/(22500 + 4147) = 0.9735
1-(2206 + 2139)/(22500 + 6717) = 0.8512
To develop a computational model for protein-protein interaction prediction, we assembled a dataset of unique proteins found as a union of those in STRING, GO for Mtb and CD. Seven features were computed for every protein-pair in the proteomes of both the organisms: Gene Ontology features for cellular component, molecular function and biological process, STRING features based on four genomic contexts, namely gene neighbourhood, gene fusion, gene co-occurrence and gene co-expression. Interactions with experimental evidence provided in STRING were used as the gold standard interacting labels. The total data collected for Mtb from the above sources consisted of 4,020 proteins and correspondingly, about 8 million protein pairs, and 3,439 known interacting pairs. CD data consisted of 4,147 proteins, about 8 million protein pairs and 3,414 known interacting pairs. It is common practice to consider randomly generated protein-pairs as non-interacting pairs for training and evaluation of PPIs (after removing any known interactions from this randomly generated set) because there is no authoritative dataset on non-interacting protein-pairs combined with the fact that only one in a thousand or so pairs is expected to interact.
For classification of interacting pairs from non-interacting pairs, we chose to use a random forest classifier as it has previously been evaluated in comparison to Bayesian network models and support vector machines, and has been shown to be most suitable for PPI prediction [32, 33]. For each bacterium, a training dataset consisting of 2,500 experimentally known interactions (positive labels) and 120,000 random pairs (negative labels) was created. A blind test set consisting of 500 positive and 150,000 random pairs is created for Mtb and CD separately. The positive to negative ratio in the test sets was chosen to mimic the real world scenario more closely where the chance that a random protein pair is interacting is as low as 1 in 500. For each scenario, a random forest consisting of 14 trees is built using the training set and 3 random features at each branch-point of a tree. Subsequent to evaluation on the test set, the random forest model was used to make predictions for each of the 8 million protein-protein pairs in both Mtb and CD. For cross species training we have used the training set developed for Mtb and evaluated the results by using the test set developed for CD.
Case I: intra-species interactome prediction
Model accuracy on evaluation dataset
The random forest model evaluates each pair and outputs a distribution value ranging between 0 and 1, where larger values correspond to a higher confidence that the pair is interacting. By setting a threshold on this distribution and taking all protein pairs above the threshold, a binary classification value is assigned to each pair. Model accuracy is computed in terms of precision (what percentage of the predicted interactions are known to be true), and recall (what percentage of known interactions are also predicted).
Predicting new interactions
What would the size of the interactome be?
Known and estimated sizes of interactomes
Number of proteins
number of pairs
Upper estimate of number of interactions
Lower estimate of number of interactions
Currently known interactions
% relative to larger estimate of interactome size
% relative to smaller estimate of interactome size
In order to evaluate the performance of our model, we carried out some analysis of the overlap of our predictions with the widely used STRING based scores as well as with the interactions determined by Wang et al using 2-hybrid technology. Figure 6 shows the overlap between these datasets (our random forest, STRING combined score, STRING experimentally known, and those by Wang et al). In Figure 6A, the overlap is shown with all pairs that have a nonzero STRING combined score. All of the random forest predictions have a nonzero STRING score. What is surprising is that for Mtb, 97% of the bacterial 2-hybrid interactions are not found among non-zero STRING combined score dataset and also have very little overlap with experimentally known interaction data of STRING, despite the fact that their method estimates that their data contains 57-61% true positives  In Figure 6B, the overlap is shown considering only those pairs that have a STRING combined score greater than 700. Corresponding numbers for CD are shown in Figures 6C and 6D. What is to be observed here is that majority of the high confidence interactions predicted by random forest and most interactions identified by Wang et al are missed by this threshold on STRING score. What is to be considered as the interactome for these bacteria? We believe that strong computational methods such as random forests can combine information from multiple sources for accurately extending the interactome beyond experimentally-known interactions.
Functional significance of interactions predicted for Mtb
The 708 high confidence interactions predicted through our method span across different functional categories. Additional file 1: Table S1 provides a detailed list of these interactions and their GO annotations of molecular function and biological process as well as the KEGG pathways that they belong to , where available. 307 of them involve proteins in the same pathway whereas 54 connect two different pathways. For 348 pairs, KEGG pathway is not known for at least one of the two proteins involved in the interaction.
The high-confidence interactions serve as hypotheses to design focussed experiments to study the role of the proteins. Among those predicted, two are described here:
Rv1538c - Rv3676: Rv1538c (ansA), L-asparaginase involved in conversion of asparagine to aspartate is predicted to be interacting with Rv3676. The protein Rv3676 is transcriptional regulatory protein and belongs to CRP/FNR family. The functional significance involving orthologs of both the proteins have been studied in Escherichia coli . cAMP receptor protein (CRP, ortholog of Rv3676) positively regulates transcription of asparaginase gene (ortholog of Rv1538c). Under the anaerobic conditions, the CRP regulatory protein play major role in facilitating the expression of asparaginase. Similar function is expected to be observed in Mtb by detail inspection of this interaction.
Rv0350 - Rv3914: Rv0350 (dnaK) is a chaperone. It is a heat shock protein and is regulated positively by Rv3223c (sigH) protein and negatively regulated by Rv0353 (hspR). This protein is predicted to be interacting with Rv3914 (trxC). The protein Rv3914 is thioredoxin and participates in various redox reactions. In Mtb, it has been observed that under the conditions of oxidative and heat stress, sigH regulates the expression of thioredoxin and dnaK .
Case II: model accuracy on cross species evaluation
Case III: model generation with ortholog transfer
To find the orthologs for CD in Mtb, we used the KEGG Orthology database . The number of orthologs found between CD and Mtb are given in Table 1, which shows the orthophylogram among Mtb, CD, yeast and human. Three datasets were made from the original CD data used in the cross-species methods. The first was the control, consisting of all CD pairs with their own GO annotations. The second dataset was made to mimic a situation in which few GO annotations were available. This was done by taking the control and removing all GO annotations for any genes with orthologs in Mtb. Of all the protein pairs used in the testing and training sets, 119,562 pairs had their GO annotations removed. The final dataset was a modification of the second set. This involved inserting the GO annotations of the orthologs in Mtb to their corresponding protein pairs in the dataset. A total of 79,331 GO annotations were transferred from Mtb to CD. This number falls short of the 119,562 values that were removed in the process of creating the second dataset. This is because the GO annotation coverage of Mtb is 71.8%, so some orthologs in Mtb are also without annotations.
In the transfer of GO annotations with orthologous genes, a consistent improvement was observed (see Figure 9). The use of ortholog transfer was between 8-10% more accurate than the data without gene annotations across the board. In the upper ranges of the model's accuracy, it managed to achieve accuracy equal to the control method.
We presented three scenarios of availability of proteome annotations for bacteria and presented approaches for rapid development of a computational system for bacterial interactome prediction, even in scenarios where little or no interaction or feature information associated with protein-pairs is available, resulting in significant savings in cost, time and effort. The methods have been demonstrated with two candidate bacterial proteomes, namely that of Mycobacterium tuberculosis (Mtb) and Clostridium difficile (CD). For the first scenario, we developed random-forest classifiers which filtered functional linkages obtained from the STRING database for Mtb and CD, by adding gene ontology features to the STRING features. We could identify potential biophysical interactions amongst these linkages with a very high precision using the classifier. In the second scenario, we demonstrated computational transfer across species by predicting CD interactions using the classifier trained over Mtb interactome. This work demonstrates that with the availability of a comprehensive interactomes for a few organisms, interctomes for other organisms may be predicted not only bioinformatically but also by transfer of computational models. We also demonstrated cross-species information may be used not only for interolog transfer, but also for computational prediction of interactions where only one of the two proteins in the pair may have an ortholog, or when both pairs may have an ortholog but not necessarily their interaction. These methods can be used for bacterial proteomes for which few features and/or interactions are available.
Features of protein-pairs used to in computational model
Gene Neighborhood score, obtained from STRING
Real valued, ranges from 0-999
Gene Fusion score, obtained from STRING
Real valued, ranges from 0-999
Gene Co-occurrence score, obtained from STRING
Real valued, ranges from 0-999
Gene Co-expression score, obtained from STRING
Real valued, ranges from 0-999
GO Cellular component based feature
Real valued, ranges from 0-0.5
GO: Molecular function
Real valued, ranges from 0-0.5
GO: Biological process
Real valued, ranges from 0-0.5
Gene ontology features
Gene Ontology (GO) provides a controlled vocabulary to describe gene products, and the relations among the terms in the vocabulary. The terms are arranged in the form of a directed acyclic graph (DAG), with three major branches from the root corresponding to molecular function (MF), cell component (CC) and biological process (BP). Annotations of gene products by these terms are created by members of the GO project by curating literature. When the DAG is traversed away from the root, the terms describe more and more specific characteristics of the gene product, whereas when traversing towards the root they become more general. For some genes only broader information is known (e.g. molecular function is hydrolase activity), whereas for some other proteins more specific information may be known (e.g. molecular function is hydrolase activity, acting on ester bonds). For use in some applications such as the current work, a smaller subset of vocabulary called GO Slims is available, which maps lower level (more specific) GO terms to a higher-level Slim category. In this work, the GO Slim annotations have been used. All GO terms associated with a protein are mapped to the GO Slim categories that were their closest ancestor in the GO DAG. For instance, Endoplasmic Reticulum (GO:0005783) is the closest ancestor of the term ER proteasome core complex, alpha-subunit complex (GO:0031605) and is also present in the GO-Slim as a broad cellular localization category. Thus GO:0031605 is mapped on to the broader cellular-localization term of Endoplasmic Reticulum. Using this approach, we assigned GO Slim categories to every protein, based on their GO term annotations. For molecular function and biological process, we made use of the GO-Slim OBO files provided by the gene ontology website, but for cellular localization, we used the GO Slim categories that we had previously created manually. The complete mapping of specific GO terms to each of the broader GO Slim terms or categories is provided for download at http://severus.dbmi.pitt.edu/TB/.
Most approaches for protein-protein interaction prediction encode the fact that a protein often shares similar features as its interacting partners. In the context of GO annotations, it is assumed that a protein would likely localize to the same cellular component destinations, or partake in the same or similar biological process, or perform comparable molecular functions as that of its interacting partners. By this reasoning often, similarity scores between GO terms of the protein-pair are encoded as a feature for PPI prediction. However, these methods do not take into consideration the possible correlations that exist between different GO categories of interacting partners. For example, it is often observed that a protein having some interacting partners in the extracellular region and some in the cytoplasm, is restricted to reside in the cell membrane, and not necessarily in the same localization as its interacting partners. Thus instead of using GO similarity as a feature, we decided to incorporate correlations between GO categories of interacting proteins in the training data, as a feature for PPI prediction.
A contingency table of term associations is computed showing number of co-occurrences of each pair of GO-slim terms that occur in known interactions (i.e. one GO slim term is associated with one protein, while the other term is associated with its interacting partner) in the training data. This is done separately for CC, MF and BP terms. From this, the feature value for a pair of proteins P i and P j having GO slim terms i and j is computed as the odds ratio (N ij /N i N j ) where Nij represents the value from contingency table corresponding to GO slim terms i and j. Ni and Nj represent individual frequency of occurrence of terms i and j amongst all annotations of all interacting proteins in the training data. This is done individually for CC, MF and BP, resulting in three features. In the use of orthologous GO annotations, the contingency tables were made from the Mtb interactome, as we would assume that had CD not been thoroughly researched, a CD contingency table would not have enough information to properly represent the co-occurrence of Gene Ontology annotations.
STRING features and interactions
STRING database provides experimentally known interactions and pairs that are predicted to have functional linkages. It assigns a score to a protein-pair indicating the functional linkages that are present between them. STRING scores derived from gene neighbourhood, gene co-occurrence, gene co-expression and gene fusion were taken as provided by the database. These form the 4 features representing a protein-pair in our dataset. Additionally, the "combined score" provided by STRING was also downloaded for comparison with other methods but is not in our model. All protein-pairs with a non-zero STRING experimentally-derived score were considered to be interacting proteins.
Orthologous genes were obtained from the KEGG Ortholog database .
The Weka package (Weka 3.6.0)'s Random Forest implementation was used to build the classifier model from the training data .
We used standard precision and recall plots to evaluate the model's performance. For varying thresholds, both precision and recall associated with the model was recorded, and a graph showing the recall on the x-axis and the precision on the y-axis was constructed. Thus, when different models are to be compared, one can do so, by observing their recall values for similar values of precision or vice-versa and see if for all values of precision, there is a general trend towards improvement in recall.
This work has been funded in part by the BRAINS grant MH094564 awarded to MG by the National Institute of Mental Health of National Institutes of Health (NIMH/NIH) of USA. AK and AG carried out the work during a summer internship at the Supercomputer Education and Research Centre (SERC) at the Indian Institute of Science; MG, AK and AG thank Prof. N. Balakrishnan for valuable scientific inputs.
- Sears CL: A dynamic partnership: celebrating our gut flora. Anaerobe. 2005, 11 (5): 247-251. 10.1016/j.anaerobe.2005.05.001.View ArticlePubMedGoogle Scholar
- Organisation WH: Multidrug and extensively drug-resistant TB (M/XDR-TB): 2010 global report on surveillance and response. 2010Google Scholar
- Koch R, Brock TD, Fred EB: The etiology of tuberculosis. Rev Infect Dis. 1982, 4 (6): 1270-1274. 10.1093/clinids/4.6.1270.View ArticlePubMedGoogle Scholar
- Walzl G, Ronacher K, Hanekom W, Scriba TJ, Zumla A: Immunological biomarkers of tuberculosis. Nat Rev Immunol. 2011, 11 (5): 343-354. 10.1038/nri2960.View ArticlePubMedGoogle Scholar
- Rachman H, Kaufmann SHE: Exploring functional genomics for the development of novel intervention strategies against tuberculosis. International Journal of Medical Microbiology. 2007, 297 (7-8): 559-567. 10.1016/j.ijmm.2007.03.003.View ArticlePubMedGoogle Scholar
- Corbett EL, Watt CJ, Walker N, Maher D, Williams BG, Raviglione MC, Dye C: The growing burden of tuberculosis: global trends and interactions with the HIV epidemic. Arch Intern Med. 2003, 163 (9): 1009-1021. 10.1001/archinte.163.9.1009.View ArticlePubMedGoogle Scholar
- Stabler RA, Gerding DN, Songer JG, Drudy D, Brazier JS, Trinh HT, Witney AA, Hinds J, Wren BW: Comparative phylogenomics of Clostridium difficile reveals clade specificity and microevolution of hypervirulent strains. J Bacteriol. 2006, 188 (20): 7297-7305. 10.1128/JB.00664-06.PubMed CentralView ArticlePubMedGoogle Scholar
- Viswanathan VK, Mallozzi MJ, Vedantam G: Clostridium difficile infection: An overview of the disease and its pathogenesis, epidemiology and interventions. Gut Microbes. 2010, 1 (4): 234-242. 10.4161/gmic.1.4.12706.PubMed CentralView ArticlePubMedGoogle Scholar
- Chandra N, Kumar D, Rao K: Systems biology of tuberculosis. Tuberculosis. 2011, 91 (5): 487-496. 10.1016/j.tube.2011.02.008. doi: 10.1016/j.tube.2011.02.008. 1472-9792View ArticlePubMedGoogle Scholar
- Raman K, Bhat AG, Chandra N: A systems perspective of host-pathogen interactions: predicting disease outcome in tuberculosis. Mol Biosyst. 2010, 6 (3): 516-530. 10.1039/b912129c.View ArticlePubMedGoogle Scholar
- Browne F, Zheng H, Wang H, Azuaje F: From Experimental Approaches to Computational Techniques: A Review on the Prediction of Protein-Protein Interactions. Advances in Artificial Intelligence. 2010, 2010: 15-924529View ArticleGoogle Scholar
- Raman K, Chandra N: Mycobacterium tuberculosis interactome analysis unravels potential pathways to drug resistance. BMC Microbiol. 2008, 8 (1): 234-10.1186/1471-2180-8-234.PubMed CentralView ArticlePubMedGoogle Scholar
- Osmanbeyoglu HU, Ganapathiraju M: Rapid development of computational prediction system of virus-host protein-protein interactions for new viruses. American Medical Informatics Association Translational Bioinformatics Summit. 2011, San Francisco, USAGoogle Scholar
- Tastan O, Qi Y, Carbonell JG, Klein-Seetharaman J: Prediction of interactions between HIV-1 and human proteins by information integration. Pac Symp Biocomput. 2009, 516-527.Google Scholar
- Venkatesan K, Rual JF, Vazquez A, Stelzl U, Lemmens I, Hirozane-Kishikawa T, Hao T, Zenkner M, Xin X, Goh KI, et al: An empirical framework for binary interactome mapping. Nat Methods. 2009, 6 (1): 83-90. 10.1038/nmeth.1280.PubMed CentralView ArticlePubMedGoogle Scholar
- Jothi R, Przytycka T: Computational approaches to predict protein-protein and domain-domain interactions. Bioinformatics Algorithms: Techniques and Applications. Edited by: Mandoiu II. 2008, Zelikovsky A: Wiley PressGoogle Scholar
- Zhong W, Sternberg PW: Genome-wide prediction of C. elegans genetic interactions. Science. 2006, 311 (5766): 1481-1484. 10.1126/science.1123287.View ArticlePubMedGoogle Scholar
- Sharan R, Suthram S, Kelley RM, Kuhn T, McCuine S, Uetz P, Sittler T, Karp RM, Ideker T: Conserved patterns of protein interaction in multiple species. Proc Natl Acad Sci USA. 2005, 102 (6): 1974-1979. 10.1073/pnas.0409522102.PubMed CentralView ArticlePubMedGoogle Scholar
- Pellegrini M, Marcotte EM, Thompson MJ, Eisenberg D, Yeates TO: Assigning protein functions by comparative genome analysis: Protein phylogenetic profiles. Proc Natl Acad Sci. 1999, 96 (8): 4285-4288. 10.1073/pnas.96.8.4285.PubMed CentralView ArticlePubMedGoogle Scholar
- Dandekar T, Snel B, Huynen M, Bork P: Conservation of gene order: a fingerprint of proteins that physically interact. Trends in Biochemical Sciences. 1998, 23 (9): 324-328. 10.1016/S0968-0004(98)01274-2.View ArticlePubMedGoogle Scholar
- Pazos F, Valencia A: In silico two-hybrid system for the selection of physically interacting protein pairs. Proteins: Structure, Function, and Bioinformatics. 2002, 47 (2): 219-227. 10.1002/prot.10074.View ArticleGoogle Scholar
- Wojcik J, Schächter V: Protein-protein interaction map inference using interacting domain profile pairs. Bioinformatics (Oxford, England). 2001, 17 (suppl 1): S296-S305. 10.1093/bioinformatics/17.suppl_1.S296.View ArticleGoogle Scholar
- Ezkurdia I, Bartoli L, Fariselli P, Casadio R, Valencia A, Tress ML: Progress and challenges in predicting protein-protein interaction sites. Briefings in Bioinformatics. 2009, 10 (3): 233-246.View ArticlePubMedGoogle Scholar
- Matthews LR, Vaglio P, Reboul J, Ge H, Davis BP, Garrels J, Vincent S, Vidal M: Identification of potential interaction networks using sequence-based searches for conserved protein-protein interactions or 'interologs'. Genome Res. 2001, 11 (12): 2120-2126. 10.1101/gr.205301.PubMed CentralView ArticlePubMedGoogle Scholar
- Liu Y, Liu N, Zhao H: Inferring protein-protein interactions through high-throughput interaction data from diverse organisms. Bioinformatics (Oxford, England). 2005, 21 (15): 3279-3285. 10.1093/bioinformatics/bti492.View ArticleGoogle Scholar
- Maggi S, Massidda O, Luzi G, Fadda D, Paolozzi L, Ghelardini P: Division protein interaction web: identification of a phylogenetically conserved common interactome between Streptococcus pneumoniae and Escherichia coli. Microbiology (Reading, England). 2008, 154 (Pt 10): 3042-3052.View ArticleGoogle Scholar
- Szklarczyk D, Franceschini A, Kuhn M, Simonovic M, Roth A, Minguez P, Doerks T, Stark M, Muller J, Bork P, et al: The STRING database in 2011: functional interaction networks of proteins, globally integrated and scored. Nucleic Acids Res. 2011, D561-D568. 39 Database
- Stark C, Breitkreutz BJ, Chatr-Aryamontri A, Boucher L, Oughtred R, Livstone MS, Nixon J, Van Auken K, Wang X, Shi X, et al: The BioGRID Interaction Database: 2011 update. Nucleic Acids Res. 2011, D698-D704. 39 Database
- Stark C, Breitkreutz BJ, Reguly T, Boucher L, Breitkreutz A, Tyers M: BioGRID: a general repository for interaction datasets. Nucleic Acids Res. 2006, D535-D539. 34 Database
- Ashburner M, Ball CA, Blake JA, Botstein D, Butler H, Cherry JM, Davis AP, Dolinski K, Dwight SS, Eppig JT, et al: Gene ontology: tool for the unification of biology. The Gene Ontology Consortium. Nat Genet. 2000, 25 (1): 25-29. 10.1038/75556.PubMed CentralView ArticlePubMedGoogle Scholar
- Aoki-Kinoshita KF, Kanehisa M: Gene annotation and pathway mapping in KEGG. Methods in molecular biology (Clifton, NJ). 2007, 396: 71-91. 10.1007/978-1-59745-515-2_6.View ArticleGoogle Scholar
- Lin N, Wu B, Jansen R, Gerstein M, Zhao H: Information assessment on predicting protein-protein interactions. BMC Bioinforma. 2004, 5: 154-10.1186/1471-2105-5-154.View ArticleGoogle Scholar
- Qi Y, Bar-Joseph Z, Klein-Seetharaman J: Evaluation of different biological data and computational classification methods for use in protein interaction prediction. Proteins. 2006, 63 (3): 490-500. 10.1002/prot.20865.PubMed CentralView ArticlePubMedGoogle Scholar
- Sambourg L, Thierry-Mieg N: New insights into protein-protein interaction data lead to increased estimates of the S. cerevisiae interactome size. BMC Bioinforma. 2010, 11: 605-10.1186/1471-2105-11-605.View ArticleGoogle Scholar
- Peri S, Navarro JD, Kristiansen TZ, Amanchy R, Surendranath V, Muthusamy B, Gandhi TK, Chandrika KN, Deshpande N, Suresh S, et al: Human protein reference database as a discovery resource for proteomics. Nucleic Acids Res. 2004, D497-D501. 32 Database
- Wang Y, Cui T, Zhang C, Yang M, Huang Y, Li W, Zhang L, Gao C, He Y, Li Y, et al: Global Protein-Protein Interaction Network in the Human Pathogen Mycobacterium tuberculosis H37Rv. Journal of Proteome Research. 2010, 9 (12): 6665-6677. 10.1021/pr100808n.View ArticlePubMedGoogle Scholar
- Kanehisa M, Goto S: KEGG: kyoto encyclopedia of genes and genomes. Nucleic Acids Res. 2000, 28 (1): 27-30. 10.1093/nar/28.1.27.PubMed CentralView ArticlePubMedGoogle Scholar
- Alifano P, Fani R, Lio P, Lazcano A, Bazzicalupo M, Carlomagno MS, Bruni CB: Histidine biosynthetic pathway and genes: structure, regulation, and evolution. Microbiol Rev. 1996, 60 (1): 44-69.PubMed CentralPubMedGoogle Scholar
- Cho Y, Ioerger TR, Sacchettini JC: Discovery of novel nitrobenzothiazole inhibitors for Mycobacterium tuberculosis ATP phosphoribosyl transferase (HisG) through virtual screening. J Med Chem. 2008, 51 (19): 5984-5992. 10.1021/jm800328v.PubMed CentralView ArticlePubMedGoogle Scholar
- Hinshelwood S, Stoker NG: Cloning of mycobacterial histidine synthesis genes by complementation of a Mycobacterium smegmatis auxotroph. Mol Microbiol. 1992, 6 (19): 2887-2895. 10.1111/j.1365-2958.1992.tb01468.x.View ArticlePubMedGoogle Scholar
- Jennings MP, Beacham IR: Co-dependent positive regulation of the ansB promoter of Escherichia coli by CRP and the FNR protein: a molecular analysis. Mol Microbiol. 1993, 9 (1): 155-164. 10.1111/j.1365-2958.1993.tb01677.x.View ArticlePubMedGoogle Scholar
- Raman S, Song T, Puyang X, Bardarov S, Jacobs WR, Husson RN: The alternative sigma factor SigH regulates major components of oxidative and heat stress responses in Mycobacterium tuberculosis. J Bacteriol. 2001, 183 (20): 6119-6125. 10.1128/JB.183.20.6119-6125.2001.PubMed CentralView ArticlePubMedGoogle Scholar
- Hall M, Frank E, Holmes G, Pfahringer B, Reutemann P, Witten IH: The WEKA Data Mining Software: An Update. SIGKDD Explorations: 2009. 2009Google Scholar
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.